lichess.org
Donate

Stockfish 16 beta

#21
Maybe we need a common FEN code to test which engine versions will give the correct move response.
We could do our own testing on our systems and then set one engine as default.

youtu.be/0GueZpYt7Oo?si=HJHWwFiwA9xiP1Pd

A lichess benchmark section would be nice to have as a reference point.
We could then know which computer works with which stockfish version.
#22
First test game with the new engines:



I wanted SF 14 NNUE to play as Black but it couldn't be loaded properly, so I used the SF 16 HCE-version, White was SF 16 NNUE 60 MB. No opening book used, all moves are results of the standard 8 sec. calculation - which didn't really work. Most of the time they calculated for about 15 sec. and about 22-30 moves deep, then stopped working. Only one variation was calculated and the winning move was played in one window, then I switched to another window and played the move there too and then let the other engine calculate the move for the other color. Just as it was suggested in the mail.

White clearly dominated Black and got an easy win. It felt like SF 16 60 MB was 200-300 Elo points higher, at least. SF 16 NNUE 60 was always at 500-800 Kn/s while SF 16 HCE sometimes went up to 1 Mn/s or even more, didn't help though.
#23
I think so like the prevision version of stockfish has said that d4 results to equality but new version says that there is a .3 advantage to white!! ..........................GOOD GOING SO FAR..............................
#26
Second test game with the new engines:



This time it was SF 16 · 7MB NNUE as White vs. SF 16 · 40MB NNUE as Black, with the same settings as before (8 sec., one var., 3/4 threads, 1 GB memory, no opening book). During the game both sides tried to avoid threefold move repetition a couple of times. Especially move 20 was interesting: SF 16 7 MB evaluated the situation as slightly better for White (+0.1) while SF 16 40 MB was at 0.0 for 10 moves already. I think that was the reason why White played Bd3 instead of going for the draw. Later in the game White found the idea of a minority attack but Black's pieces were ideally placed to stop the b4-b5 pawn push. Until queens and a rook were exchanged and in move 127 b5 was played. But with only a knight and a rook left it was more or less a forced threefold repetition. Interesting game, but lots of pieces shuffling which is quite common for these engine games.

I think I've found out how the "Search time" mode actually works: Let's say you choose "8 sec." and the engine calculates for up to 7 seconds and reaches depth 30. One would think that the search would end one second later, but that is not the case. Both engines worked until they went up to depth 31 - no matter if that takes only one or even thirty seconds. Since SF 16 7 MB always had higher depth (and also higher kn/s - Mn/s) it altogether calculated a lot longer than the 40 MB version, which had a far lower search depth and also a lot less kn/s.
#27
Depth search is not available.
If you select Search time infinity we can wait till we see the depth we want to stop the search at.
Feature request, offer a variable search depth setting, so that it stops on the depth we asked it to stop.
Every move would then have the same depth search.
#28
The graph is showing Stockfish 12 finding mates better than SF16.
github.com/vondele/matetrack

Adjusting optimal local settings like, Threads (80% max threads), Hash (More memory is best), MultiPV vs PV1 only for best move ...
github.com/official-stockfish/Stockfish/wiki/Stockfish-FAQ#interpretation-of-the-stockfish-evaluation

Monitor the CPU temperature when running a chess engine:
You might want to calculate loads within a 20 % safety margin. Example: If you have 4 threads on your PC, use 3 or less for your chess engine. If you selected 3Threads/2Core, than close other programs to ensure accurate results. Overloading is never a good thing.

Thermal throttling = reduced performance: If you hear your fan speeding up and down, then at least check the temperature of your CPU. You do not want it to over heat. If it happens, use one less thread for your analysis until it stops hunting.

Using windows run command: Win+R
Type> perfmon.exe /res

Take note of what you PC is running at, without engine running and then compare with engine running.
Example: Before ...
CPU 2% CPU Usage 61% Maximum Frequency
Memory 0 Hard Faults/sec 52% Used Physical Memery

Engine Analyse running ... SF 11 HCE (3T/2C)
CPU 81% CPU Usage 98% Maximum Frequency
Disk 32 KB/sec Disk I/O 1% Highest Activity Time
Memory 0 Hard Faults/sec 62% Used Physical Memory

HWiNFO: I select "Sensors only" ...
www.hwinfo.com/download/

Join the Lichess Beta Testers team, to post in this forum